Fabien Niñoles
@ninoles.bsky.social
Dancer, poet, RPG theorist, FOSS enthusiast and developer. Cofounder & CTO https://genvid.com. Earthling.
Why mechanistic interpretability is important? By better understanding how a LM works, we can fix some of his bias, better train them, and more importantly, built more efficient solution that do the same things more deterministically.
2/2
2/2
November 19, 2025 at 6:06 PM
Why mechanistic interpretability is important? By better understanding how a LM works, we can fix some of his bias, better train them, and more importantly, built more efficient solution that do the same things more deterministically.
2/2
2/2
Yes, sorry. Given my engineering background, I did find the distinction between black box and grey box important. But yes, for most people it is the same.
For me, the difference means you can't do mechanistic interpretability on a black box.
1/..
For me, the difference means you can't do mechanistic interpretability on a black box.
1/..
November 19, 2025 at 6:06 PM
Yes, sorry. Given my engineering background, I did find the distinction between black box and grey box important. But yes, for most people it is the same.
For me, the difference means you can't do mechanistic interpretability on a black box.
1/..
For me, the difference means you can't do mechanistic interpretability on a black box.
1/..
Darn, maybe we should avoid trying to find a single generic approach for all technical problems?
November 19, 2025 at 5:44 PM
Darn, maybe we should avoid trying to find a single generic approach for all technical problems?
It's more of a grey box: open but incomprehensible. Here again an excellent video from Welch Lab:
youtu.be/UZDiGooFs54?...
youtu.be/UZDiGooFs54?...
The moment we stopped understanding AI [AlexNet]
YouTube video by Welch Labs
youtu.be
November 19, 2025 at 5:41 PM
It's more of a grey box: open but incomprehensible. Here again an excellent video from Welch Lab:
youtu.be/UZDiGooFs54?...
youtu.be/UZDiGooFs54?...
There is a paper about the costs and the limits to achieve those rates. Welch Lab has a good video about it:
youtu.be/5eqRuVp65eY?...
youtu.be/5eqRuVp65eY?...
AI can't cross this line and we don't know why.
YouTube video by Welch Labs
youtu.be
November 19, 2025 at 5:36 PM
There is a paper about the costs and the limits to achieve those rates. Welch Lab has a good video about it:
youtu.be/5eqRuVp65eY?...
youtu.be/5eqRuVp65eY?...
Yes, they seem to have improved it since I tried a few weeks ago. Idem for Gemini and Claude. Still, you never know when it will fail, and that's the main problem. And sure, humans are not absolved from errors and typo, but its why you asked for sources. LLM are doing a poor job there, yet.
November 19, 2025 at 4:43 PM
Yes, they seem to have improved it since I tried a few weeks ago. Idem for Gemini and Claude. Still, you never know when it will fail, and that's the main problem. And sure, humans are not absolved from errors and typo, but its why you asked for sources. LLM are doing a poor job there, yet.
A few weeks ago, all popular chats were giving me such non-sense answer. They seem to have improved on that, but in general rule, if you're looking for precise information, like dates, position, or other kind of identity information, you have a non-negligeable chance they have something wrong.
November 19, 2025 at 4:40 PM
A few weeks ago, all popular chats were giving me such non-sense answer. They seem to have improved on that, but in general rule, if you're looking for precise information, like dates, position, or other kind of identity information, you have a non-negligeable chance they have something wrong.
Reposted by Fabien Niñoles
4. the disregard for the corrosive power of anthropomorphism, which is taken advantage of by industry to sell & steal our data, in the base case scenario, and in the worst to abuse and push vulnerable groups to dependance and worse.
(Section 3.4 here doi.org/10.5281/zeno...)
6/n
(Section 3.4 here doi.org/10.5281/zeno...)
6/n
September 6, 2025 at 8:34 AM
4. the disregard for the corrosive power of anthropomorphism, which is taken advantage of by industry to sell & steal our data, in the base case scenario, and in the worst to abuse and push vulnerable groups to dependance and worse.
(Section 3.4 here doi.org/10.5281/zeno...)
6/n
(Section 3.4 here doi.org/10.5281/zeno...)
6/n
No need to go that far: here Mistral's answer for the age of Susan Calvin:
November 19, 2025 at 4:29 PM
No need to go that far: here Mistral's answer for the age of Susan Calvin:
Most of current LLM based deployments lack a post-generation validation mechanism. A good example of such system is Alpha Geometry, which use LLM to generate constructs, and a symbolic engine to validate.
deepmind.google/blog/alphage...
deepmind.google/blog/alphage...
AlphaGeometry: An Olympiad-level AI system for geometry
Our AI system surpasses the state-of-the-art approach for geometry problems, advancing AI reasoning in mathematics
deepmind.google
November 19, 2025 at 3:41 PM
Most of current LLM based deployments lack a post-generation validation mechanism. A good example of such system is Alpha Geometry, which use LLM to generate constructs, and a symbolic engine to validate.
deepmind.google/blog/alphage...
deepmind.google/blog/alphage...
Even with clean data, LLM (like any ANN) will still push some garbage out. That is basically what training is about: reduce (not eliminate!) the amount of garbage out through gradient descent on error.
The surprising thing about LLM is how often it is right. The garbage is the expected part.
The surprising thing about LLM is how often it is right. The garbage is the expected part.
November 19, 2025 at 3:25 PM
Even with clean data, LLM (like any ANN) will still push some garbage out. That is basically what training is about: reduce (not eliminate!) the amount of garbage out through gradient descent on error.
The surprising thing about LLM is how often it is right. The garbage is the expected part.
The surprising thing about LLM is how often it is right. The garbage is the expected part.
For me, LLM's deception is not about its lies, but the lies around it:
* anthropomorphize interface and terminology
* no understanding of how the outcome is reach
* impossible to came with a confidence factor on the results
* cost paid != resources consumed
* very undeserved hype around it
* anthropomorphize interface and terminology
* no understanding of how the outcome is reach
* impossible to came with a confidence factor on the results
* cost paid != resources consumed
* very undeserved hype around it
November 19, 2025 at 3:19 PM
For me, LLM's deception is not about its lies, but the lies around it:
* anthropomorphize interface and terminology
* no understanding of how the outcome is reach
* impossible to came with a confidence factor on the results
* cost paid != resources consumed
* very undeserved hype around it
* anthropomorphize interface and terminology
* no understanding of how the outcome is reach
* impossible to came with a confidence factor on the results
* cost paid != resources consumed
* very undeserved hype around it
I think what he means is that LLM doesn't even know how to reproduce lies or deceipts. They can generate sentences that no humans have ever said, like "There are 5 shapes in this image, which are a square, a circle and a triangle." That sentence is just wrong and it doesn't know that.
November 19, 2025 at 2:58 PM
I think what he means is that LLM doesn't even know how to reproduce lies or deceipts. They can generate sentences that no humans have ever said, like "There are 5 shapes in this image, which are a square, a circle and a triangle." That sentence is just wrong and it doesn't know that.
Ask LLM what are the states with a r in their name and see if it makes consensus. Sounds like a trick question? Well, now, tell me, how do you distinguish those cases from the others? When can you tell that it got it wrong? LLM's hallucinations are never consensus. They are blatant inventions.
November 19, 2025 at 2:41 PM
Ask LLM what are the states with a r in their name and see if it makes consensus. Sounds like a trick question? Well, now, tell me, how do you distinguish those cases from the others? When can you tell that it got it wrong? LLM's hallucinations are never consensus. They are blatant inventions.
Ajoutez à ça un NPD complètement décharné et qui n'a pas encore fini de lécher ses blessures et ce n'est pas pour rien que seulement le BQ s'est opposé concrètement.
C'est malheureux mais bien le résultat de l'influence polarisante de la politique américaine sur notre système politique.
2/2
C'est malheureux mais bien le résultat de l'influence polarisante de la politique américaine sur notre système politique.
2/2
November 19, 2025 at 2:04 PM
Ajoutez à ça un NPD complètement décharné et qui n'a pas encore fini de lécher ses blessures et ce n'est pas pour rien que seulement le BQ s'est opposé concrètement.
C'est malheureux mais bien le résultat de l'influence polarisante de la politique américaine sur notre système politique.
2/2
C'est malheureux mais bien le résultat de l'influence polarisante de la politique américaine sur notre système politique.
2/2
L'article est bizarre: l'auteure reproché au PLC son "arrogance" mais la cause sur la désorganisation de l'opposition. À mon avis, le problème est plus du côté du CPC qui est devenu sous Harper plus réactionnaire que conservateur et donc incapable de former une opposition effective.
1/..
1/..
November 19, 2025 at 2:04 PM
L'article est bizarre: l'auteure reproché au PLC son "arrogance" mais la cause sur la désorganisation de l'opposition. À mon avis, le problème est plus du côté du CPC qui est devenu sous Harper plus réactionnaire que conservateur et donc incapable de former une opposition effective.
1/..
1/..
Anti-terrorist laws are always rooted into a desire for authoritarianism from the government in place (including the minority parties). It's a failure of democracy, of governance, saying that our military and judicial systems cannot succeed without reducing the rights of the bystanders.
November 19, 2025 at 1:34 PM
Anti-terrorist laws are always rooted into a desire for authoritarianism from the government in place (including the minority parties). It's a failure of democracy, of governance, saying that our military and judicial systems cannot succeed without reducing the rights of the bystanders.
The game has 20 endings achievable in exactly 31 days, and 3 of them are considered victories. I think you can only achieve one of them by approving everything (if you answer correctly at some final questions).
November 19, 2025 at 1:06 PM
The game has 20 endings achievable in exactly 31 days, and 3 of them are considered victories. I think you can only achieve one of them by approving everything (if you answer correctly at some final questions).
That's impossible. He thinks humanity exists for him and sees himself as Nietzsche's übermensch (perverted), Humanity's Ultime Finality (not what Nietzsche means). So, everything good for him must be good for humanity and the only things that are good for humanity are things that are good for him.
November 19, 2025 at 4:43 AM
That's impossible. He thinks humanity exists for him and sees himself as Nietzsche's übermensch (perverted), Humanity's Ultime Finality (not what Nietzsche means). So, everything good for him must be good for humanity and the only things that are good for humanity are things that are good for him.
He insulted women and others quite often, and treated them like objects even those few he likes. But for me, it was the imperative "Quite, Piggy!" which tries to establish a feudal dominance ("I'm your lord, you're a lowly serf") that really makes it more upsetting.
November 19, 2025 at 2:23 AM
He insulted women and others quite often, and treated them like objects even those few he likes. But for me, it was the imperative "Quite, Piggy!" which tries to establish a feudal dominance ("I'm your lord, you're a lowly serf") that really makes it more upsetting.
That happens when people think those methods are just different tools to do development the "old" way.
Iterative development (not just agile) is a different approach and most agile methodologies are tools built upon them.
But you must do ID first for them to make sense.
Iterative development (not just agile) is a different approach and most agile methodologies are tools built upon them.
But you must do ID first for them to make sense.
November 19, 2025 at 2:12 AM
That happens when people think those methods are just different tools to do development the "old" way.
Iterative development (not just agile) is a different approach and most agile methodologies are tools built upon them.
But you must do ID first for them to make sense.
Iterative development (not just agile) is a different approach and most agile methodologies are tools built upon them.
But you must do ID first for them to make sense.
De plus, je reprends en mes mots ce qu'il dit plus éloquemment que moi, mais c'est loin d'être une citation. Je ne prétends donc pas que c'est de lui, bien que c'est assez proche de ce qu'il dit.
November 18, 2025 at 8:59 AM
De plus, je reprends en mes mots ce qu'il dit plus éloquemment que moi, mais c'est loin d'être une citation. Je ne prétends donc pas que c'est de lui, bien que c'est assez proche de ce qu'il dit.
Désolé, c'est tellement connu cette citation, je n'ai pas pensé qu'il était nécessaire de le préciser.
Un peu comme parler de la relativité sans mentionner Einstein.
Maintenant, je crois que ce serait bien que plus de monde s'approprie cette pensée.
Un peu comme parler de la relativité sans mentionner Einstein.
Maintenant, je crois que ce serait bien que plus de monde s'approprie cette pensée.
November 18, 2025 at 8:54 AM
Désolé, c'est tellement connu cette citation, je n'ai pas pensé qu'il était nécessaire de le préciser.
Un peu comme parler de la relativité sans mentionner Einstein.
Maintenant, je crois que ce serait bien que plus de monde s'approprie cette pensée.
Un peu comme parler de la relativité sans mentionner Einstein.
Maintenant, je crois que ce serait bien que plus de monde s'approprie cette pensée.
Your cat would have been the first to remind you if you had forgotten. That's just what cats expect.
November 17, 2025 at 9:38 PM
Your cat would have been the first to remind you if you had forgotten. That's just what cats expect.